Research On Action Recognition Method Based On Feature Representation

Action recognition is an important field in the computer vision area, and has been wid-ly applied to many vision-based systems, such as smart surveillance, video retrieval, video summarization, intelligent robot, human-computer interaction, etc. Similar to common com-puter vision problems such as object detection and recognition, the key of action recognition method is feature representation of video. Due to the enormous variations of natural scene video in background, view, scale, illumination, etc, and the diversity of human action on the object appearance and execution, feature representation of video is challenging and can greatly affect the performance of action recognition method.Feature representation methods map video from sample space to feature space. Ac-cording to the amount of semantics captured by features, the feature representation methods can be simply divided into three catgories:low-level, mid-level and high-level. In this dissertation, action recognition methods based on mid-level and high-level feature represen-tation are studied on the basis of analysis and summarization of the existing video feature representation methods.Firstly, a hierarchical feature graph model is proposed based on analyzing the rela-tionships among low-level, mid-level and high-level feature representations. The proposed model obtains a high-level feature representation from low-level features by constructing feature graphs layer by layer. The feature graph in each layer integrates the feature contents with spatiotemporal relationships between features to describe action completely. To rec-ognize actiones by the proposed model, the method of building feature graph of each layer is concretely described, and also a hierarchical graph matching algorithm is introduced to compute the similarity between two videos. Experimental results on several public action datasets demonstarte the effectiveness.Secondly, a method for action recognition based on automatical discovery of action parts is proposed in consideration of insufficient representative power of high-level features and better representative power and discriminative power of mid-level features. The pro- posed method discovers action parts by trianing candidate action part detectors and select-ing distinctive action part detectors. For the former, the feature whitening is incorporated with cross validation to ensure the coherence of each candidate detector. For the latter, a novel Coverage-Entropy Curve is proposed to evaluate the capability of detector, and the corresponding similarity measurement is also defined to remove redundancy of detectors. Experimental results on four public action datasets show that the action parts discovered by the proposed method can classify actions effectively.Thirdly, a method for automatically discovering action parts based on feature selection is proposed to address the problem of selecting effective part detectors. Common selection algorithms adopt heuristic rules, which fail to analyze the capability of classifing actions of part detector, and can’t ensure that the selected detectors could recognize actions optimally. Therefore, we convert the problem of selecting effective part detectors to a standard feature selection problem, and propose two kinds of solving methods, i.e. SVM based algorithm and sparse representation based algorithm from different perspective. We conduct experiments on four public action datasets, and our method outperforms the state-of-the-art.At last, the presented work is summarized. According to the imperfect aspects, future work has also been discussed.